DexPour: Effective and Efficient High-DoF Robotic Hand Liquid Pouring via Hierarchical Reward with Approximated Proxy Abstraction

Abstract

Pouring fluids is a routine task for humans but challenging for high-DoF robots, particularly given fluid simulation’s computational demands while training policies. In this paper, we propose DexPour, a novel reinforcement learning method with hierarchical rewards and Approximated Proxy Abstraction (APA) method. APA efficiently approximates liquid behavior using a small set of spheres, reducing computational overhead. Meanwhile, our hierarchical reward framework breaks down the intricate pouring process into four distinct stages—approach, grasp, transport, and pour—providing fine-grained feedback and fostering stable policy learning. Extensive experiments demonstrate that DexPour achieves a 92% fluid transfer efficiency with a 70% cup fill and a 99% efficiency at 30% fill, highlighting its robust performance across varying liquid volumes. Ablation studies highlight the contribution of each component, confirming the necessity of detailed stage-wise guidance for complex dexterous manipulation. In addition, we compare DexPour with a full fluid simulation baseline, showing comparable pouring efficiency while reducing training time by 81.6%, demonstrating DexPour’s efficiency and practical viability for fluid manipulation tasks.